BRAW? So 3:1 compression is RAW????

There is a third category between lossless and lossy that BMD has exploited and that is visually lossless. Visually lossless is still a form of lossy but it has a subjective view in that in most viewing situations of motion video, the viewer should not be aware of any loss of information.

With CinemaDNG, BMD considered 3:1 and 4:1 to be visually lossless and would not compress data further. With BRAW, BMD considers 12:1 visually lossless. Of course there’s a lot of information discarded but you may not be aware of it on practical viewing if the video.

Q0 and Q5 are also visually lossless and although the Q5 can throw away more data than 12:1 in some situations, it can also keep as much data or more than 8:1.

Everyone here appreciates pixel peeping, in which we have proven you can see the loss when comparing apples and oranges but the results of video in motion may still be acceptable to most viewers.

Yes I’m with Rick on this one. It’s definitely lossy, and you can see it at q5 and 12:1 when cropping. It’s not crazy bad or anything, but it can be noticeable. 8:1 seem to be the borderline. It’s quite nice balanced between datarate and IQ, almost appearing lossless. Is not, but there is less obvious sign of data lost.

If you look at pictures of braw and CDNG side by side, you will see BRAW with less detail. However, it’s harder to tell looking at a film clip in motion. BRAW is foremost a motion picture codec, not still photo. Cdng on the other spawn from DNG which primary is for photo, giving it another characteristic. BRAW appears much better when it comes to motion. Not more detail, but more fluid as of original film.

Good comparison, Oyvind. Of course DNG was designed for stills and BRAW is all about video. CinemaDNG lossless and lossy are not the same.

I think with a maximum bitrate that represents a 2:1 compression, BRAW Q0 is the best bridge to lossless since we can’t have compressed BRAW lossless. For any complex material of shorter duration or very important work like a commercial, go Q0. After that maybe 5:1 for general purpose, and then consider Q5 especially for high volume work (documentaries or low budget long narratives). That would cover the range of extremes in three steps.

Still early days trying to work out what is best.


Sent from my iPhone using Tapatalk

Video is just set of still displayed one after another(with bit of motion blur).

Yes, but perception is different from looking at stills. If you'd have ever looked at a single still from a classic movie you've seen in cinema, you'd be surprised how grainy and soft it seems. Our brain is integrating detail from consecutive frames, and that can play a role with digital compression too.

A typical example where this doesn't work any more are the drone shots with too much small detail for the compression, where we can perceive the GOPs as pulsating sharpness. A really good GOP codec with enough data rate doesn't do that to perceived detail.

I think a good compression algorithm for film is taking perception of consecutive frames into consideration, and in my impression BRAW does (even if it's not a GOP codec). Plus, please keep in mind, that many say BRAW does some filtering and reduces aliasing to some degree. DNG doesn't filter at all, and as long as you use a BM camera without an OLPF, DNG will give an unrealistic impression of sharpness by some false detail (for false detail, read about the Shannon/Nyquist limit). The difference between DNG and BRAW is getting much smaller if you have an OLPF.

It's lossy. I don't know if it ever gets to true visually lossless. I'll lay it out. About a decade ago real lossless might he up to 4:1. But Braw is JPEG 1 from what we know, so with good processing 2:1-3:1. But what do the braw ratios mean. If they are comparing to actual Bayer frames, then they are substantially higher compression in real life, as they are encoded into larger non bayer frames. So, is 12:1 really 18:1+ compression. Meaning, 3:1 would be 4.5:1+, not one mode would be true lossless. You see why I say Braw should move to JPEG xs base, which I think has a Bayer mode.

Now, is braw visually lossless. If so, the bank note scene would still have the faint coloured lines on the note. Making something visible disappear, and clearly visible at that, is hardly visually lossless. Also, the scene with the netting/mesh, where it just goes to a blurred out patch, is not visually lossless. Where is that s going crop up and kill the image, actresses with bright blond hair

I remember the days when we had 300 bit per second phone coupled dial up modems, it feels kind of like that. Let's hope that Bayer patent dies, so we can get back to better quality. Sure Braw might remove a lot of false detail, along with genuine detail, but you can do better than that in processing cdng, because it was genuine detail as the sensor presented to the codec. Full data detail, and you can't get better than that at that level. But cdng is still jpeg1.

Let's look at what future Braw based on true Bayer compression, and Jpeg xs Bayer compression looks like. One, 3:1-4:1 lossless would be my guess. 6:1+ visually lossless to a genuine extent. 6:1average variable true lossless might be possible.

If you write lossless codec at guaranteed 4:1 for every recording then you can make millions. So far anything lossless is around/below 2:1.
It may be possible only after quite strong denoising, which always kills details as well (hello BRAW).
Instead of this write efficient Bayer optimised 3:1 codec. This is already good enough for every case if codec is good.

Of course denoise, the noise isn't inn the scene, its in peoples head

On average 4:1 should be possible. Not every frame, and maybe not every shoot. But I've discussed things with David and maybe others, and back then virtually nothing got to 3:1, a lot in the 2.x range, with JPEG lower. But compression also depends on the compression techniques and that has gone ahead in the last 12+ years. I remember back then finding a lossless compressor that did 6:1, but denoised simple images in 4:4:4, high bit depth like 12 bit+, all the easy stuff to compress.


My personal technological aim is 9:1 to over 100:1. The latter is through inter techniques. But this requires understanding everything, and people don't do that.

Denoising doesn't have to reduce any more detail then what's there. Temporal. Not what Braw is doing, is a very bad way.

You won't get 4:1 with sensor data (probably even with acceptable denoise). Ffv1 is probably most complex and ratio orientated and no way it will do 4:1 on never compressed footage. Of course at this point you also start to have problems with encoding and decoding speeds ( ffv1 is not crazy fast). With 8k you have to be realistic and in this case 2:1 is about it.
For me more sense is to do good, RAW orientated codec at 3:1. Then you know you ratio so can plan storage etc.

Yes, it is slow, given. But faster circuit technology, no problem eventually. So don't discount as one day it will be possible.

However, I disagree, if you can do proper noise removal, "proper" that is, it should be possible given good settings. The following settings give more compressible image and is higher quality image acquisition. 4:4:4, 16 bit, higher frame rate. The following is even better, as they are for computational imaging where you can set the lens look latter in software. Near infinite depth of field, sorry the others just slipped my mind because of brain issue. More than one way to skin a cat.

Proper noise removal I think is a big stride. But also smart technique, where normal compression of this type is blunt I think.

Now, yes 3:1, but I think applying the above 6:1-12:1 real visually lossless is worth aiming for as well.